This seems like a shot across the bow for all large Claude API customers, which I'm sure they saw coming.
But still, a TSMC style pure play model provider would win huge business in the space given how many application companies are being eaten by model companies.
I'm surprised more people are not talking about the fact that the two best models in the world, Gemini 3 and Claude 4.5 Opus, were both trained on Google TPU clusters.
Presumably, inference can be done on TPUs, Nvidia chips, in Anthropic's case, new stuff like Trainium.
I despise Facebook and all that it stands for, but if the surplus value that it has extracted from humanity over the last two decades is reinvested intelligently into nuclear energy, I'm actually okay with it.
Despite the hype that you see on Twitter, the hard tech startup scene is actually incapable of large-scale engineering coordination on the level needed for a nuclear power plant, or even a gas turbine.
If any innovation on fission reactors is going to be successfully commercialized, we will need to see billions of dollars of investment over medium to long time horizons.
Of course, the millstone around the neck of nuclear power is that it's a dual-use technology. There's probably a lot more behind the scenes that's been done to stifle the industry effectively for non-proliferation reasons, but masquerading as cost, regulatory problems, environmental concerns, etc.
I would buy a Cybertruck tomorrow if it had a gas engine. I would buy a $10,000 or $15,000 gas generator add-on if it enabled unlimited range (provided I have gasoline).
There are just too many places, even in California, where I have to limit my trip because of electric range.
Given recent high profile redaction events, I think one simple use of AI would be to have it redact documents according to an objective standard.
That should in theory prevent overly redacted documents for political purposes.
An approach that could be rolled out today would be redacting with human review, but showing what % of redactions the AI would have done, and also showing the prompt given to the AI to perform redactions.
Whether or not the patent was actually granted in this case, I have not been able to think of a compelling reason to have patents for software. In fact, I think most intellectual property laws need to be seriously rethought.
If the objective is to maximize investment by protecting successful results, I don't think our system is doing a very good job.
Who can do this with good data controls? I don't want to have to dig through the fine print of some Terms of Service page to figure out if a sequencing company is going to save a copy of my genetic code for possible future use.
I don’t think Rust syntax and patterns (no classes) are especially elegant for many tasks. I can’t express the behavior of a system as cleanly in Rust as TypeScript, C#, go or Python. I know that’s not what it was designed for, but a guy can dream.
But what Rust has is the best tooling bar none(cargo, build system, compile time checks, ease of first use). The tooling is actually more important than the borrow checker and memory safety in my opinion.
If I clone a Rust repo, it’s actually easier to compile, test, and run the code than any other language. It avoided the fragmentation of JS/TS (npm, yarn, pnpm, bun, deno) and dep hell of Python (which was a nightmare until uv).
If Rust didn’t have the syntax warts (macros), it would be eating the world.
The idea that artifacts belong forever to whoever inhabits the land today is going to put under increasing pressure as ancient DNA continues to reveal the number and severity of population replacements over time.
Unless there is a trick that I am missing, I don't think this will work by itself. The fundamental thing is what can the model attend to as it generates the next token.
If you give a summary+graph to the model, it can still only attend to the summary for token 1. If it's going to call a tool for a deeper memory, it still only gets the summary when it makes the decision on what to call.
You get the same problem when asking the model to make changes in even medium-sized code bases. It starts from scratch each time, takes forever to read a bunch of files, and sometimes it reads the right stuff, other times it doesn't.
There's a chance this memory problem is not going to be that easy to solve. It's true context lengths have gotten much longer, but all context is not created equal.
There's like a significant loss of model sharpness as context goes over 100K. Sometimes earlier, sometimes later. Even using context windows to their maximum extent today, the models are not always especially nuanced over the long ctx. I compact after 100K tokens.